Archives

ALS GeoAnalytics needed to scale machine learning model training and inference for core logging analysis while managing computational costs effectively.

Synthesia needed to maximize GPU utilization during video inference on EC2 G7e instances by reducing idle time caused by sequential GPU compute, data transfer, and post-processing operations.

Determining whether security-focused LLMs can effectively identify vulnerabilities in live production infrastructure code at scale.

Enabling engineers to run multiple concurrent coding sessions and integrating AI agents into automated internal workflows at scale.

Enable on-device AI models to coordinate complex tasks across external data sources while maintaining persistent user context and proactive engagement without relying solely on cloud connectivity.

AI agents needed a standardized way to generate UI components that work across different platforms and frameworks without being tightly coupled to any specific technology stack.

Enabling efficient execution of generative AI models on edge devices with limited computational resources while maintaining acceptable latency and performance.

Developers needed a way to build AI agent workflows that could run on Android devices and backend systems without reinventing the core agentic logic across different platforms.

Developers need a way to reliably control, monitor, and extend AI model generation calls in production agentic applications without modifying core business logic.

Running large language models efficiently on mobile and edge devices while preserving multimodal and agentic capabilities without requiring server-side inference.

Building production-grade AI agents that can maintain context and state across long-running enterprise workflows spanning days or weeks without losing information during idle periods or server restarts.

Mobile developers faced performance and battery inefficiency when running AI models on CPU/GPU, limiting real-time AI applications on edge devices.

Developers needed a unified embedding model capable of processing interleaved multimodal inputs (text, images, video, audio, documents) in a single semantic space for tasks like retrieval-augmented generation and visual search.

How can Google enable third-party service providers and hardware manufacturers to build intelligent smart home experiences without requiring deep AI/ML expertise or significant R&D investment?

Developers needed a unified way to build, deploy, and run high-performance machine learning models directly on edge devices (Google Pixel TPU) with reliable fallback mechanisms.

Enabling efficient post-training of large language models on single-host TPU configurations without requiring complex multi-host distributed setups.

Developers needed accessible infrastructure, resources, and structured learning pathways to effectively build and optimize AI applications using GPUs and large language models at scale.

Converting a brittle, monolithic sales research AI prototype into a production-ready agent that eliminates silent failures, fragile parsing, and lacks observability.

AI training pipelines were bottlenecked by slow data I/O when accessing training datasets stored in Google Cloud, limiting throughput and increasing total training time.

Autoregressive LLM decoding suffers from sequential bottlenecks where tokens must be generated one-at-a-time, limiting throughput and inference speed on hardware accelerators like TPUs.

How to deploy high-intelligence AI models with agentic capabilities to consumer hardware and mobile devices without requiring cloud infrastructure.

Enterprise systems need to react to events in real-time rather than relying on slow batch jobs or inefficient polling microservices that create dangerous delays in detecting critical issues like fraud or supply chain disruptions.

Organizations need to securely build, deploy, and govern autonomous AI agents at enterprise scale as the industry transitions from experimental LLMs to production agentic AI systems.

Automating the transformation of raw community signals into reliable technical guidance at scale using multiple specialized agents.

Deploying and managing AI agents at scale in production requires infrastructure for state management, security governance, and complex workflow orchestration that goes beyond demo implementations.

How to enable developers to build multimodal AI agents that can process and respond to real-time audio, video, text, and generation capabilities beyond traditional text-based interfaces.

BASF needed to manage and optimize thousands of interdependent supply chain decisions across 180 global production sites where weather and regulatory changes can cause cascading disruptions in a two-year production pipeline.

Building safe, reliable, and autonomous agents that can act independently across multiple enterprise systems while maintaining security, governance, and reliability guardrails.

AI agents built on Google Cloud need access to accurate, current, and grounded information about Google's products and APIs to function effectively.

Google needed to accelerate large-scale codebase migrations (TensorFlow to JAX) that are too complex and interconnected for manual developer effort or standard AI coding tools to handle efficiently.

Developers needed a unified, secure way to build AI agents locally and deploy them to Google Cloud with standardized protocols and tooling.

Efficiently evaluating and validating LLM-generated outputs at scale during experimentation without manual review bottlenecks.

Netflix needed to manage the lifecycle of machine learning models across multiple domains and teams at scale, moving beyond their original single-domain personalization focus.

Netflix needed to automatically evaluate the quality and relevance of show synopses at scale to improve member discovery and engagement.

Netflix needed to efficiently extract and surface key moments from hundreds or thousands of hours of raw video footage for editorial teams to accelerate the creative content production process.

Netflix needed to design a domain-independent traffic routing system for their ML model serving infrastructure that could handle personalized experiences at scale across multiple domains while maintaining high availability.

How to identify and surface the most interesting and meaningful listening moments from a year's worth of user streaming data to create personalized narrative highlights for Wrapped.

Spotify needed to optimize ad targeting and delivery at scale by coordinating multiple specialized systems to make smarter advertising decisions rather than relying on monolithic ad selection logic.

Detecting and preventing first-party fraud at scale across a payment network where legitimate users abuse policies through multiple accounts, free trial cycling, and refund exploitation.

Detecting and preventing fraudulent behavior in free trial signups, such as repeated trial abuse and missed cancellations, at scale with high accuracy.

Detecting and preventing sophisticated fraud attacks while minimizing friction for legitimate users in payment systems.

Traditional rule-based KYC (Know Your Customer) systems lack the autonomous decision-making capability and real-time validation speed needed for modern financial services compliance operations.

Cloudflare needed to scale code review processes across their engineering organization while maintaining code quality and security standards without overwhelming human reviewers.

Cloudflare needed to build an internal AI engineering stack that could handle massive scale (20 million requests, 241 billion tokens) while dogfooding their own platform products.

Facebook Groups Search was unreliable at helping users discover and validate community content most relevant to their search queries.

Providing a scalable, efficient search infrastructure that allows AI agents to dynamically create search instances and perform semantic queries across uploaded documents without managing underlying indexing complexity.

AI agents lack persistent memory mechanisms to retain context, learn from interactions, and improve decision-making over time.

AI agents needed a way to interact with browsers at scale while maintaining visibility and control over automated actions, requiring higher concurrency and real-time debugging capabilities.

How to efficiently run inference for extra-large language models on edge infrastructure while maintaining low latency and high throughput across distributed Cloudflare servers.

Developers needed a unified way to access multiple AI model providers without managing separate integrations and API contracts for each one.

Building a scalable platform for deploying AI agents at the edge that can think, act, and persist state across distributed Cloudflare infrastructure.

GPU memory bandwidth constraints were limiting LLM inference efficiency across Cloudflare's distributed edge network, requiring optimization to deliver faster and cheaper inference.

Meta needed to automatically identify and remediate performance inefficiencies across their massive infrastructure to reduce power consumption and free up engineering capacity.

Simplifying the deployment and scheduling of machine learning inference workloads across multiple instances and instance types on Amazon SageMaker HyperPod.

AI coding assistants were ineffective at making useful edits in large-scale data pipelines because they lacked sufficient understanding of complex, multi-repository codebases spanning multiple languages and thousands of files.

AI agents struggle to iterate rapidly on system design and codebases due to architectural patterns that limit their ability to understand, modify, and validate applications effectively.

Detecting safety hazards in real-time across hundreds of distributed operational sites using video feeds while maintaining low latency and managing the computational complexity of processing multiple camera streams.

Aigen needed to scale machine learning pipelines across hundreds of distributed edge solar robots while managing data labeling and model training challenges in agricultural robotics.

Building forecasting models that remain accurate during sudden market shocks like a global pandemic, where historical data no longer predicts future outcomes.

Detecting sophisticated client-side security threats like zero-day exploits while minimizing false positives in real-time across millions of requests.

How to safely execute untrusted AI-generated code with minimal latency and resource overhead.

Designing high-quality, sustainable concrete mixes that are produced in the United States while optimizing for performance characteristics.

Meta needed to automatically optimize low-level infrastructure and kernel-level parameters for AI ranking models to improve performance without manual tuning.

Meta needed to scale their ads ranking models to LLM-scale complexity and size while maintaining inference latency requirements for real-time ad serving.

Training and evaluating AI models is resource-intensive, requiring significant human effort to generate quality training data and assess model outputs.

Advancing AI research requires collaboration between industry and academia, but funding and partnership models need structured programs.

Data science teams need diverse skill sets that blend mathematical rigor with creative problem-solving to build effective ML systems.

LinkedIn's Feed needed to evolve to handle increasing content diversity, real-time ranking signals, and personalization at massive scale.

LinkedIn's LLM-based ranking systems faced latency and throughput challenges when serving personalized results at scale.

Building personalized generative AI features at LinkedIn's scale required a robust and reliable application infrastructure that could serve millions of users.

Responding to operational events in Amazon EKS clusters is often manual, slow, and requires deep expertise, making it difficult to handle incidents at scale across complex Kubernetes environments.

Organizations building generative AI workloads on AWS lacked comprehensive architectural guidance covering responsible AI, data architecture, and emerging patterns like agentic workflows, leading to poorly architected AI systems.

Organizations building ML workloads on AWS lacked up-to-date architectural guidance that incorporates the latest services, capabilities, and best practices, leading to sub-optimal ML system designs across reliability, performance, cost, and operational dimensions.

Diagnosing and resolving issues in complex Kubernetes clusters is slow and requires expert knowledge, leading to high Mean Time to Recovery (MTTR) and heavy reliance on specialized engineers for root cause analysis.

Organizations deploying AI/ML workloads on AWS lacked comprehensive architectural guidance for building responsible, well-architected machine learning and generative AI systems at scale.

Enterprises adopting Amazon Bedrock need centralized governance over AI model access, including authorization controls, usage quotas, and auditing, but lack a standardized gateway pattern to enforce these policies at scale.

Artera needed to develop and scale an AI-powered prostate cancer diagnostic test, requiring significant compute resources for model training/inference and a reliable pipeline to deliver timely, personalized treatment recommendations.

Airbnb needed to advance its AI, data science, and machine learning capabilities across multiple domains (NLP, optimization, measurement science) to improve its travel and living platform, requiring solutions to challenges in search ranking, recommendation, experimentation, and large-scale data processing.

Producing valid and realistic mock data for GraphQL testing and prototyping is tedious to write and maintain; existing approaches like random value generation and field-level stubbing lack domain context, resulting in unconvincing and brittle test data that doesn't scale across a large schema.

Airbnb needed to build robust data science and economic modeling capabilities to understand and optimize their two-sided marketplace dynamics for policy and business decisions.

Airbnb users in the early trip planning stage often lack a clear travel destination, making it difficult to provide relevant recommendations and convert exploratory browsing into bookings.

Organizations struggle to discover and secure AI-powered applications across their infrastructure, especially shadow AI deployments that teams spin up without central oversight, creating security blind spots.

Running large AI models for agent workloads on edge infrastructure was cost-prohibitive and required significant inference stack optimization to serve models like Kimi K2.5 efficiently at scale.

AI agents hitting Cloudflare error pages received heavyweight HTML responses that consumed excessive tokens and required brittle parsing, making automated error handling inefficient and costly.

Enterprise search and AI assistant products like Dropbox Dash need to connect disparate data sources and optimize AI-driven retrieval, but naively querying across siloed data with LLMs leads to poor relevance and brittle prompt engineering.

Large machine learning models require significant memory and compute resources, making deployment and inference expensive and slow, especially in resource-constrained environments.

Dropbox Dash's AI agent struggled with effectiveness when naively providing all available context to the model, leading to degraded performance as irrelevant information diluted the signal needed for accurate, agentic AI responses.

Running AI inference for products like Dropbox Dash at scale is expensive and resource-intensive, requiring efficient use of compute and memory to make the product accessible to a broad user base.

Manual prompt engineering for Dropbox Dash's relevance judge was unreliable, hard to measure, and costly—making it difficult to systematically improve task performance in production.

Dropbox Dash needs to rank and retrieve relevant context across a user's work in real time, requiring low-latency access to precomputed and real-time features for AI-driven search and recommendation models.

Engineering organizations face open questions about how to effectively integrate AI coding tools (like Claude Code and Cursor) into developer workflows and where these tools can have the most measurable impact on productivity.

Dash's search ranking models required large volumes of high-quality labeled relevance data to train effectively, but human labeling alone was too slow and expensive to scale to the needed coverage.

Dropbox Dash needed deeper understanding of multimodal content (photos and videos) across user files, but processing diverse media types at Dropbox's scale posed efficiency and architectural challenges.

Connecting thousands of GPUs across multiple data centers and regions for gigawatt-scale AI training clusters requires seamlessly bridging different network fabrics, which creates massive networking and interconnect challenges.

Facebook Reels needed a way to enhance social discovery by surfacing content that friends have interacted with, requiring real-time computation of relationship strength and ranking of friend-engaged content at massive scale.

Updating security-related APIs across millions of lines of code and thousands of engineers is extremely difficult at scale, especially when a single class of mobile vulnerability can be replicated across hundreds of locations in an Android codebase.

GPU-to-GPU communication performance on AMD platforms was insufficient for Meta's evolving AI model training workloads, and the standard RCCL library didn't meet the performance and flexibility requirements of their internal workloads.

Meta's ads ranking ML experimentation lifecycle required extensive manual intervention from engineers for hypothesis generation, training job launches, failure debugging, and result iteration, slowing down the pace of ranking model innovation.

Agentic (AI-driven) software development produces and ships code so fast that traditional testing frameworks cannot keep pace, leaving bugs uncaught as they land in rapidly evolving codebases.

Netflix needed scalable, deep machine-level understanding of every piece of content across an expanding catalog (including live events and podcasts) to power recommendations and discovery, but building separate models per content type and modality doesn't scale.

Netflix's Ranker service had a video serendipity scoring feature (computing how different a title is from a user's watch history) consuming ~7.5% of total CPU per node, creating a significant performance bottleneck at their enormous scale.

Generic pre-trained LLMs lack the domain-specific alignment needed for Netflix's production use cases in recommendation, personalization, and search, and the post-training pipeline to fine-tune them doesn't scale efficiently across multiple domain constraints and reliability requirements.

Netflix's Graph Search platform for federated enterprise data required users to write structured queries, limiting accessibility and ease of use despite the system being scalable and configurable.